Update README.md
Browse files
README.md
CHANGED
|
@@ -545,35 +545,39 @@ This repo contains several multilingual parallel benchmarks covering European la
|
|
| 545 |
|
| 546 |
## Collection
|
| 547 |
As of **October, 2026**, it includes
|
| 548 |
-
- MMLU
|
| 549 |
- `cais/mmlu`
|
| 550 |
- `openai/mmmlu`
|
| 551 |
- `CohereLabs/Global-MMLU`
|
| 552 |
-
- `edinburgh-dawg/mmlu-redux`
|
| 553 |
- `collections/INSAIT-Institute/multilingual-benchmarks`
|
| 554 |
- `Eurolingua/mmlux`
|
| 555 |
-
- BELEBELE
|
| 556 |
-
- ARC-Easy / ARC_Challenge
|
| 557 |
-
- PolyMath
|
| 558 |
-
|
| 559 |
-
|
| 560 |
-
Sources: lukfre/mmlu-mix (MMLU-Redux-EU), , allenai/ai2_arc + Eurolingua/arcx, Qwen/PolyMath; licences as in each source.
|
| 561 |
|
| 562 |
|
| 563 |
## Schema
|
| 564 |
Every multiple-choice benchmark gets the same schema, so one set of `lm-eval` templates covers all:
|
| 565 |
|
| 566 |
-
```
|
| 567 |
-
|
| 568 |
-
|
| 569 |
question question text (the passage is separate, in `passage`, for belebele)
|
| 570 |
choices option texts, already in the NEW order
|
| 571 |
answer index of the correct option in `choices` (NEW order)
|
| 572 |
answer_original index of the correct option in the source order
|
| 573 |
perm perm[new_position] = source position, so choices[i] == source_choices[perm[i]]
|
| 574 |
shuffled False when the item kept its source order (see "Pinned options")
|
| 575 |
-
|
| 576 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 577 |
```
|
| 578 |
|
| 579 |
## Shuffle rules
|
|
|
|
| 545 |
|
| 546 |
## Collection
|
| 547 |
As of **October, 2026**, it includes
|
| 548 |
+
- MMLU-mix, our own build, cleaned and refined from:
|
| 549 |
- `cais/mmlu`
|
| 550 |
- `openai/mmmlu`
|
| 551 |
- `CohereLabs/Global-MMLU`
|
| 552 |
+
- `edinburgh-dawg/mmlu-redux-2`
|
| 553 |
- `collections/INSAIT-Institute/multilingual-benchmarks`
|
| 554 |
- `Eurolingua/mmlux`
|
| 555 |
+
- BELEBELE, copied from `facebook/belebele`
|
| 556 |
+
- ARC-Easy / ARC_Challenge, copied from `allenai/ai2_arc` + `Eurolingua/arcx`
|
| 557 |
+
- PolyMath, copied from `Qwen/PolyMath`
|
|
|
|
|
|
|
|
|
|
| 558 |
|
| 559 |
|
| 560 |
## Schema
|
| 561 |
Every multiple-choice benchmark gets the same schema, so one set of `lm-eval` templates covers all:
|
| 562 |
|
| 563 |
+
```output
|
| 564 |
+
<data>
|
| 565 |
+
id item key shared by all languages (mmlu "subject/idx", belebele "link#qnum", ARC id)
|
| 566 |
question question text (the passage is separate, in `passage`, for belebele)
|
| 567 |
choices option texts, already in the NEW order
|
| 568 |
answer index of the correct option in `choices` (NEW order)
|
| 569 |
answer_original index of the correct option in the source order
|
| 570 |
perm perm[new_position] = source position, so choices[i] == source_choices[perm[i]]
|
| 571 |
shuffled False when the item kept its source order (see "Pinned options")
|
| 572 |
+
|
| 573 |
+
<metadata>
|
| 574 |
+
lang
|
| 575 |
+
translation_tier
|
| 576 |
+
source_dataset
|
| 577 |
+
|
| 578 |
+
<custom>
|
| 579 |
+
+ mmlu: subject, subject_category, idx
|
| 580 |
+
+ belebele: passage
|
| 581 |
```
|
| 582 |
|
| 583 |
## Shuffle rules
|